ValueGen Class — pytorch Architecture

Source Code

aten/src/ATen/test/vec_test_all_types.h lines 609–652

template <typename T>
struct ValueGen<T, true, false>
{
    std::mt19937 gen;
    std::normal_distribution<reduced_fp_to_float_t<T>> normal;
    std::uniform_int_distribution<int> roundChance;
    T _start;
    T _stop;
    bool use_sign_change = false;
    bool use_round = true;
    ValueGen() : ValueGen(std::numeric_limits<T>::min(), std::numeric_limits<T>::max())
    {
    }
    ValueGen(uint64_t seed) : ValueGen(std::numeric_limits<T>::min(), std::numeric_limits<T>::max(), seed)
    {
    }
    ValueGen(T start, T stop, uint64_t seed = TestSeed())
    {
        gen = std::mt19937(seed);
        T mean = start * static_cast<T>(0.5) + stop * static_cast<T>(0.5);
        //make it  normal +-3sigma
        T divRange = static_cast<T>(6.0);
        T stdev = std::abs(stop / divRange - start / divRange);
        normal = std::normal_distribution<reduced_fp_to_float_t<T>>{ mean, stdev };
        // in real its hard to get rounded value
        // so we will force it by  uniform chance
        roundChance = std::uniform_int_distribution<int>(0, 5);
        _start = start;
        _stop = stop;
    }
    T get()
    {
        T a = normal(gen);
        //make rounded value ,too
        auto rChoice = roundChance(gen);
        if (rChoice == 1)
            a = std::round(a);
        if (a < _start)
            return nextafter(_start, _stop);
        if (a >= _stop)
            return nextafter(_stop, _start);
        return a;
    }
};